Hyperspectral imaging for chlorophyll detection: 1D-2D-COS feature optimization and adaptive deep learning modeling

The chlorophyll of tomato leaves can be characterized in response to the influence of salt on leaf membrane permeability and plant growth. This study investigates the characterization of tomato leaf chlorophyll in response to salt stress by leveraging near-infrared (NIR) hyperspectral imaging (HSI). A novel predictive model integrating two-dimensional correlation spectroscopy (2D-COS) with a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) architecture was developed. To enhance model robustness, the sparrow search algorithm (SSA) was employed for parameter optimization alongside a self-attention mechanism. 2D-COS analysis identified chlorophyll-sensitive peaks at 1438 nm, 1627 nm, 1660 nm, and 993 nm, revealing the sequence of molecular bond variations. Characteristic wavelengths were selected using competitive adaptive reweighted sampling (CARS), the successive projections algorithm (SPA), and a combined one-dimensional-two-dimensional correlation spectroscopy (1D-2D-COS) approach. The SSA-CNN-BiLSTM-Multi-head-Attention model based on characteristic wavelengths extracted by 1D-2D-COS method worked best, reaching 0.9817 and 0.9178 for R C and R P , and 1.6091 and 3.1546 for RMSEC and RMSEP. The results demonstrate that the integration of 1D-2D-COS with the advanced deep learning algorithm provides an effective tool for monitoring leaf chlorophyll, offering theoretical support for the early detection of salt stress in plants. • 2D-COS combined with CNN-BiLSTM was used to predict chlorophyll content. • 2D-COS was used to locate sensitive peaks and resolve the order of chemical bonds. • 1D spectra and 2D-COS (1D-2D-CNN) were combined to extract the feature wavelengths. • SSA and self-attention mechanism was used to improve the robustness of models. • 1D-2D-COS-SSA-CNN-BiLSTM-Mutilhead-Attention model yielded the best results.